/extract-signal

An elastic methodology for turning a noisy source surface into a vetted signal in the form the requester asked for, without contaminating the context that must stay clean. It is a goal + a toolkit + the judgment to compose techniques to fit the task — not a fixed pipeline. The work it does is determined by the context it is invoked with.

Operational owner: this SKILL.md. protocol.yon is its aligned declarative companion. When either changes, update the other and refresh the @STAMP date.

The principle that motivates this skill

Context hygiene comes from RELOCATION, not TRUNCATION. Noisy work happens where noise is cheap — a spawned sub-agent's window, or this skill's own disposable window when it is itself a sub-agent. Only the signal reaches the place that must stay clean. The gathering area is never shrunk to fit an arbitrary size.

Inference & contracts (read the call, don't demand a form)

The skill infers its parameters from how it was invoked; explicit overrides are accepted but never required:

• What signal — the question to answer.
• What form — the output shape (see menu). Inferred from the ask ("as a risk-list", "is it ready?", "what are the options").
• Stakes — inferred from consequence; drives vetting depth.
• Mode & output — inferred per below.

Modes & output (inferred)

• Agentic (default — the normal call). The root agent owns composition, gathering dispatch and review dispatch. Invoke selected prime-* skills at the root, so their workers are leaves. For a cold opinion, the root dispatches a separate sibling reviewer with the question, primary-source pointers and artifact, without the gatherer's conclusions or lead rationale. Gatherers and reviewers never orchestrate each other. The root integrates and returns the signal inline unless a file is specified.
• In-session / new-session (it IS the agent). extract-signal gathers and vets directly as the running agent; output inline by default, to a file if specified. (When it is itself a dispatched sub-agent, this is the same path: it works inline in its disposable window and does NOT spawn further sub-agents — depth-1 guardrail; the isolation already protects the caller.) In a leaf, use the bounded direct-gather procedure from its primed brief, not a prime-* invocation that requires another worker. Run permitted vetting inline with nested delegation suppressed. Disclose same-context dependence; it is not a cold opinion. If material vetting requires an independent opinion, return findings and that unmet requirement to the root before a readiness verdict.

The methodology (elastic — compose, don't follow blindly)

1. Frame the inferred signal request (what · form · stakes).
2. Compose the technique set for this task and run it (in the mode above):
   • Gather — /prime-sweep (many sources) · /prime-fetch (one noisy call) · /prime-expand (vague request → questions + named sources first).
   • Diverge — /insight-explore when the signal needs options/hypotheses generated before they can be vetted.
   • Vet — /insight-critique (poke holes) → /insight-assess (evidence solid? what does it imply?).
   • Cold opinion — a fresh, no-prior-framing read (/cold-review-style) as a drift guard.
3. Vet proportionally to stakes — cheap/low-stakes → a quick critique may suffice; material/high-stakes → full critique + assess + cold opinion. Reach for the cold opinion whenever the lead has been deep in the material.
4. Present the signal in the requested form (next two sections govern how).

Output contract — the signal carries its provenance

Whatever the form, the presentation MUST include:

• the signal in the requested form;
• provenance — source:line (or equivalent) for load-bearing claims;
• confidence + residual gaps (what wasn't determined and why);
• which techniques ran (so the caller can judge the vetting depth).

Sensitivity vs destination (do not over-expose the signal)

Signal inherits the sensitivity of what it was extracted from. Before presenting:

• Tag the signal's sensitivity (e.g. internal / restricted / public).
• The destination must be cleared for that sensitivity. If the output is bound for a lower-trust or external/public destination than the signal warrants, flag it and withhold the sensitive parts rather than present them. Never let internal detail flow into a public-bound artifact by default.

Failure & contradiction handling

• Empty gather → report "no signal found", with what was searched. Never fabricate signal to fill the form.
• Vet disagreement (/insight-assess vs the cold opinion) → surface the split with both positions; do not silently average to a false consensus.
• Mid-flight blocker — if extraction hits a decision only the caller/handler can make, surface it and pause that line; do not guess past it.

Bounds

• Concurrency — the root owns at most 3 parallel leaf workers, depth-1. Every composed technique inherits that limit; an active mode owns dispatch rather than adding another layer.
• Termination — before work, bind finite tool-call, wave and elapsed-time limits to the invocation. Before dispatch also bind monitoring, cancellation and terminal success/failure/timeout outcomes. Without those controls, stop the dependent agentic route. Stop at the first exhausted bound or no-new-signal wave; retain partial findings and name capped or unperformed vetting. Do not retry or extend automatically.

Signal forms (reference menu — any custom form allowed)

Examples (illustrative — the lead composes to fit the actual call)

1 · Agentic + high-stakes → bill-of-health

Call: extract-signal: is @acme/parser ready to ship publicly? Infers: signal = release-readiness · form = bill-of-health · stakes = high · mode = agentic (default) · output = inline to caller. Composes: root invokes prime-sweep to dispatch leaf gathering → root applies inline critique and assess → root dispatches a separate cold sibling reviewer → root returns the bill-of-health. Returns: verdict + per-area findings with source:line, confidence, residual gaps, techniques-run.

2 · In-session inline + options → ranked-options

Call (running as the session): extract-signal: what are our options for the cache layer? inline. Infers: form = ranked-options · stakes = moderate · mode = in-session (it IS the agent) · output = inline. Composes: insight-explore (3+ options) → insight-assess (score) → present ranked. Lighter vet — no cold opinion at moderate stakes. Returns: ranked options + rationale + provenance + gaps.

3 · Output-to-file + moderate → risk-list

Call: extract-signal: risks in this diff → write to risks.md Infers: form = risk-list · output = file (risks.md) · stakes = moderate · mode = agentic. Composes: prime-fetch the diff → insight-critique (hunt risks) → write a severity-ranked risk-list to the file. Writes: risks.md — each risk with severity + source:line + confidence.

4 · Sensitivity-edge (the guard fires)

Call: extract-signal: summarize the auth module for the public README Infers: signal sensitivity = internal · destination = public README. Acts: sensitivity (internal) > destination (public) → withhold internal detail, flag it — present only the public-safe surface and tell the caller what was withheld and why. Never auto-leaks internal detail into a public artifact.

Rules

• MUST infer the signal request from the call; MUST NOT demand a rigid input form.
• MUST default to root-owned agentic gathering and sibling review; when it IS the agent, work inline with output inline-by-default or to a file if specified.
• MUST keep gathered output complete (report-to-file for large surfaces); MUST NOT impose an arbitrary word/size cap on gathering.
• MUST compose techniques to fit the task; MUST NOT run a fixed sequence regardless of the signal.
• MUST vet before presenting; depth MUST scale with stakes.
• MUST attach provenance + confidence + gaps + techniques-run to the presented signal.
• MUST sensitivity-match the signal to its destination; MUST NOT over-expose internal signal to a public/external destination.
• MUST surface (not average) vet disagreements and (not guess past) mid-flight blockers.
• MUST NOT spawn sub-agents from within a sub-agent (depth-1 guardrail).

Pairs with

• /prime-sweep, /prime-fetch, /prime-expand — gather techniques.
• /insight-explore — divergent generation when candidates are needed first.
• /insight-critique, /insight-assess — the vetting engine.
• /cold-review — the second-cold-opinion shape.
• /multi-agent-mode — the session posture this runs inside.
• /caller-options — surfaces the venue/output/form choices.

Evolution (logged for v2+)

A signal ledger/cache (reuse prior extractions) and signal-quality metrics are deferred — not yet in scope.

Human output. This skill's handler-facing output obeys the human-output contract (human-output/SKILL.md).

Self-improvement. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): apply its signal taxonomy and threshold to this run; surface any surviving proposal for handler approval. Conservative — silent otherwise. Never auto-apply.